Based On The Object-oriented High-resolution Uav Image Disasters And Key Technology Of Information Extraction

After the "5.12" devastating earthquake in WenChun, the geological disaster has been being in a frequent trend in Sichuan Province. The unmanned aircraft at low altitude remote sensing played an important role in the scientific and technological disaster relief. While in the face of amount of UAV high resolution Images, how to extract the information effectively has been the key to UAV low-altitude remote sensing research techniques.This paper focused on the center alleviate of UAV high resolution image disaster information extraction. We took the UAV high resolution images in Sichuan regions as the main experimental data. In order to overcome the lack of traditional information extraction methods, object-oriented approach was adopted to deal with the classification of images, and some key technologies were researched:First, grounded on the summary of previous multi-scale segmentation algorithms, multi-scale color image segmentation algorithm based on continuous texture was proposed. In the algorithm, first over-segmentation image was got, texture continuity regions were calculated fast acquired through setting texture factor as thresholds. And object regions were described by characteristics of spectral and shape. In the end, merging costs were calculated, which were limited by setting scale factors in order to get multi-scale results.Second, on the basis of analysis of each factor impact to the segmentation results. in the segmentation algorithm, quantitative selection of optimal scale method on local variance. And optimal segmentation result was got by the grid method. Our segmentation algorithm was compared with other two segmentation algorithm by visual description and contrast of region number, and we also compared it to existing main commercial software segmentation results through quantitative analysis. From the results, it demonstrated the advantages of our segmentation algorithm.Third, image objects were got quantitatively. Spectral, shape and texture features were extracted and stored-The necessity of muti-feature extraction was proved by a water extraction experiment. Then principal component analysis and other processing were implemented to the features.Fourth, object-oriented classification based on support vector machine was researched. Appropriate kernel function was selected and parameter optimization was processed. Supervised classification was done after choosing Training objects. Then the classification results were compared with object-oriented classification based on BP neural network and pixel classification based on support vector machine. The results show the classification accuracy indicators of object-oriented classification based on support vector machine was better than other two methods.Finally, Dammed lake disaster information extraction was conducted; form the Japan disaster areas images before and after the disaster, we classified the two images separately. Then disaster area was got and specific disaster information was got from comparative analysis.